DeepSeek AI Unveils DeepSeek-V4.1-Flash Featuring 1M Context and Advanced Caching

DeepSeek AI has introduced DeepSeek-V4.1-Flash, a cutting-edge multimodal model that addresses the challenges posed by long-horizon agents in large language model (LLM) serving. This model features a substantial 1M-token context window and a global key-value (KV) cache footprint of just 890 bytes per token, significantly reducing memory demands compared to its predecessors. The architecture incorporates a 40-layer structure, divided into a 20-layer causal encoder and a 20-layer decoder, which optimizes prefill computations by leveraging the final encoder hidden state. The model employs Compressed Sparse Attention 2 (CSA2), which enhances efficiency by allowing layers to share KV caches in various modes. Additionally, the FP4 KV cache is quantized for improved storage efficiency. Deployment is straightforward, with open weights available under an MIT license and a public API that supports various reasoning tiers. The training process utilized a vast dataset of multimodal tokens, achieving impressive performance metrics across several benchmarks, outperforming notable models like Opus-5 and GPT-5.6.